Research map / Image generation / Super-resolution
Rethinking Imbalance in Image Super-Resolution for Efficient Inference
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes WBSR, a plug-and-play weight-balancing framework for image super-resolution combining hierarchical equalization sampling, a balanced diversity loss, and gradient-projection dynamic inference to reduce compute (~34%) while matching or beating SR baselines.
Problem
Standard SR training samples patches uniformly and optimizes all weights equally with L1/L2 losses, ignoring data imbalance (easy flat vs. hard textured patches) and parameter redundancy, leading to overfitting in smooth regions and underfitting in textured ones.
Approach
Reframes SR as imbalanced distribution transfer learning: resamples patches by gradient-based difficulty classes (HES), replaces L1/L2 with a GMM-based balanced diversity loss (BDLoss), and routes patches to supernet subnets by gradient magnitude at inference.
Key results
- WBSR on RCAN at x4 reaches Test2K 26.45 dB at 57% FLOPs versus RCAN 26.39 dB at 100% FLOPs.
- HES outperforms prior sampling method BSPA by about 0.1 dB average PSNR.
- Abstract reports comparable or superior performance with roughly 34% reduction in computational cost.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply WBSR to RCAN | RCAN · Test2K ×4 | PSNR: 26.39 -> 26.45 | Helped |
| Apply WBSR to RCAN | RCAN · Test2K ×4 | #FLOPs (G): 32.60 (100%) -> 18.52 (57%) | Helped |
| Apply WBSR to RCAN | RCAN · Test4K ×4 | PSNR: 27.89 -> 27.94 | Helped |
| Apply WBSR to RCAN | RCAN · Test4K ×4 | #FLOPs (G): 32.60 (100%) -> 19.40( 59%) | Helped |
| Add HES sampling (ablation) | SRResNet · Test2K ×4 | PSNR: 26.19 -> 26.24 | Helped |
| Add HES sampling (ablation) | SRResNet · Test4K ×4 | PSNR: 27.65 -> 27.71 | Helped |
| Replace L1 loss with Lbd (ablation) | SRResNet · Test2K ×4 | PSNR: 26.19 -> 26.21 | Helped |
| Replace L1 loss with Lbd (ablation) | SRResNet · Test4K ×4 | PSNR: 27.65 -> 27.70 | Helped |
| Jointly use HES and Lbd with dynamic inference (ablation) | SRResNet · Test2K ×4 | PSNR: 26.19 -> 26.26 | Helped |
| Jointly use HES and Lbd with dynamic inference (ablation) | SRResNet · Test4K ×4 | PSNR: 27.65 -> 27.73 | Helped |
| Use full network for inference rather than dynamic subnets (ablation) | SRResNet + WBSR · Test2K ×4 | PSNR: 26.26 -> 26.38 | Helped |
| Use full network for inference rather than dynamic subnets (ablation) | SRResNet + WBSR · Test4K ×4 | PSNR: 27.73 -> 27.80 | Helped |
Surprising findings
- Full-network WBSR (no dynamic routing) gains about 0.33 dB PSNR over baselines at comparable FLOPs, indicating the balancing training alone helps.
Limitations
- No error bars or statistical significance reported.
- Fixed training epochs cause extra subnets to be under-trained as subnet count grows (performance drops with more subnets).
Experimental setup
Datasets: DIV2K
Benchmarks: B100 Urban100 Test2K Test4K
Metrics: PSNR SSIM FLOPs
Baselines: SRResNet RCAN ClassSR ARM BSPA SamplingAug DDA
Builds on: SRResNet RCAN ClassSR ARM
Closest papers by meaning
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.86)
- Pixel-level and Semantic-level Adjustable Super-resolution: A Dual-LoRA Approach (CVPR 2025 · similarity 0.85)
- Lightweight Image Super-Resolution via Flexible Meta Pruning (ICML 2024 · similarity 0.85)
- TSD-SR: One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution (CVPR 2025 · similarity 0.85)
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · similarity 0.85)
- Learning Coupled Dictionaries from Unpaired Data for Image Super-Resolution (CVPR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.